Specialization in Global Value Chains
From factor abundance to fragmentation
University of Groningen
Faculty of Economics and Business
Master Thesis International Economics and Business
Student Name: Anne Mostert Student ID: S1785478
E-mail: a.mostert@student.rug.nl
Abstract
The rise of global value chains has profoundly changed the landscape of modern trade giving more importance to intermediate goods as opposed to final goods. In relation to this important change this paper investigates the relationship of a fundamental outcome of traditional trade theory that trade leads to specialization. The paper draws on the methodology of input-output analysis on the basis of the recently published World Input-Output Database (WIOD). The outcome of this research gives some indicative evidence of the existence of this relationship, however this relationship seems limited only to European countries.
1
1: Introduction
Some 200 years ago David Ricardo (1817/1911)1 published his most influential work On the Principles of Political Economy and Taxation. In this book Ricardo has laid the foundation for trade theory by giving his rationale for trade using the example of trading British cloth for Portuguese wine. The revolutionary aspect of his economic thinking is that according to him the driving forces of trade are not to be found in absolute advantages in the production process but rather in the relative ones. The rationale behind his thinking is that trade causes specialization in production of those goods in which a country has a comparative advantage and this specialization will cause both countries to be better off in the end.
The world of Ricardo, however, is long past and the economic reality of today is very different from what it was 200 years ago. We no longer trade wine for cloth. The production process has become strongly fragmented because trade and communication costs have declined. It is no longer necessary for a single good to be produced in one country. Rather, parts and components of goods often come from all over the world, a phenomenon that we call global value chains. We define a global value chain as the full set of international activities that are required to produce a single good or service. The idea behind the rise of global value chains is that an ever declining amount of goods and services are produced in a single country. Assembly work nowadays is mostly done in low wage countries, and many helpdesks are now located in Bangalore rather than in the home country. A famous study by Feenstra (1998) has shown how a single Matel Barbie doll is designed in the US, the materials come from Taiwan and Japan, assembly is done in low wage countries such as Indonesia, Malaysia and China, and the dolls are sold worldwide. The rise of global value chains (or GVCs), however, is not only limited to low wage countries. For example, Apple’s iPad and iPhone are assembled in China, but also the EU, Japan, Taiwan and South Korea are heavily involved in the creation of an American iPhone or iPad by supplying intermediate products (Kraemer, Linden and Dedrick, 2010).
As will become clear, the global value chain used to involve only one, or very few countries. Today more and more countries are involved in the global value chain and in that sense global value chains have become longer. This ‘lengthening’ of the global value chain has been named many things: (international) fragmentation, off-shoring, slicing up the value chain or vertical specialization. Whatever term used, this process has changed trade, our modern globalized economy trades more intermediate rather than final goods.
In the Ricardian model trade leads to specialization, the question however remains whether
1 The original publication of Principles of Political Economy and Taxation was published in 1817, the 1911 edition is in
2 this assertion is true in our modern globalized economy, a world in which we no longer trade wine for cloth, but rather computer chips for LCD screens. This is where this paper comes in. It intends to test one outcome of traditional trade theory, namely: trade leads to specialization, also in the context of the new phenomenon of global value chain trade.
The nature of global value chain trade however makes empirical analysis much more difficult. It is no longer one country that produces a certain brand of car, the production of cars involves complex networks of different suppliers who themselves also may be involved in a complex network of suppliers and so on and so forth. We will try to analyse these linkages with the use of input-output analysis (Leontief, 1936) using the newly published World Input-Output Database (WIOD). This new database contains harmonized data on the inputs used by sectors around the world. This as we will argue is ideal for analysing the different linkages in global value chain trade.
The aim of this study is to link the rise of global value chain trade with export specialization. The concrete research question of this study is: Is there a positive link between the extent of international fragmentation and export specialization between the years 1995 and 2007? The period in question is chosen for the availability of the data, the WIOD is a recent publication and only has data going back to 1995. Although there is more recent data than 2007 we will not include this data into our analysis to avoid possible interference of the recent economic and financial crisis. It is conceivable that some countries in some sectors are struck harder by the financial and economic crisis than others. The financial sector, for instance, has been struck very hard in both the USA and in Europe due to their interconnectedness. The downfall of this sector might influence the specialization of different countries, but this effect has nothing to do with the extent to which a country participates in global value chains.
This study is subdivided into the following sections. In section 2 we will discuss some studies relevant to our research. In section 3 we will discuss the methodology and data sources we intend to use to answer the research question. We will present and discuss the results from our analysis in section 4. Finally, in section 5 we will present our conclusion and prospects for future research.
2. Literature
The literature section of this paper will be subdivided into three parts. (1) Traditional trade theory and why it leads to specialization. (2) The phenomenon of global value chain trade. (3) The theoretical linkage between fragmentation and specialization.
2.1. Traditional trade theory
3 work of Ricardo (1911) and, Hecksher (1919) and Ohlin (1933). Ricardian trade theory is very important for the development of trade theories because Ricardo was the first to recognize the benefits of trade. In the Ricardian model, countries differ in their technological capabilities. The Ricardian model assumes two goods and two countries with just one factor of production: labour. Because countries differ in their technological capabilities a country (say Portugal) will produce one good (say wine) relatively more efficiently in relation to the other good ( say cloth), in this case Portugal is said to have a comparative advantage in producing wine. Another country (say the UK) on the other hand might produce cloth relatively more efficient, therefore the UK has a comparative advantage in the production of cloth. If trade costs are higher than the relative differences in production costs the Ricardian model predicts that countries will not trade and produce both goods at home (since it consumes both goods).If, however, trade is costless, the outcome of the Ricardian model is that both countries will completely specialize in the production of a single good. Therefore lifting trade barriers will cause countries to specialize their production.
The theory of comparative advantage was extended in the Hecksher (1919) Ohlin (1933) model (H-O model) by including capital (besides labour) as a factor of production. Additionally the H-O model (in contrast to the Ricardian model) assumes no technological differences. We assume two countries producing two goods. Both goods require labour and capital to be produced, but the goods differ in the intensity it uses the factors of production. Because countries also differ in factor endowment. One country (say the UK) has relatively more capital and hence is able to produce goods that require more capital more cheaply (say Cloth). The other country (say Portugal) has more labour and hence is able to produce the labour intensive good more cheaply (say wine). Consequently countries with more capital will produce and export more capital intensive goods, while countries with more labour export more labour intensive goods. Although the explanatory mechanism differ, both the Ricardian and the H-O model predict the same thing: trade leads to specialization.
2.2. Global Value Chain Trade
Ricardo and Hecksher-Ohlin deal with trade in final goods, in trading wine for cloth for example. Many economists however, have pointed out that the production process has become increasingly more fragmented. Production of goods now occurs in a large global value chain rather than in a single location. This has radical implications for trade, much trade is no longer in final goods but rather in intermediate goods.
4 unbundling is characterised by the separation between the locations of production and consumption. Due to such developments as container shipping, the transport costs have dropped radically. This drop in transport costs meant that it no longer was the case that transport costs impeded the location of production to be different from where it is consumed. The second unbundling can be characterised by the separation of the location of the different stages of the production process itself. Whereas the first unbundling is attributed to the drop in transport costs due to technological developments such as container shipping. The second unbundling is attributed to the drop in costs of information and communication technology due to technological development. The invention of, for example, the telephone (either cellular or landline) and (broadband) internet allowed for a radical drop in communication costs. This drop in communication cost, in turn, lowered the coordination costs that are involved in having multiple locations for different stages in the production process. The process of unbundling is illustrated in figure 1.
According to Grossman and Rossi-Hansberg (2008), the fragmentation of the production process requires us to look at trade in a radically different way. The production of goods used to be confined to a single location and in the framework of trade theory we therefore where not interested in what the production process actually entails. Production of goods could be interpreted as black boxes with a whole lot going on that we do not see. There are a lot of tasks involved in the creation of a single product but since the production of this good takes place in the same location we focus on the end product rather than on the tasks involved. In the second unbundling these boxes began to open and we began to see that indeed there is a whole lot going on in this production process. According to Grossman and Rosi-Hansberg (2008), the best way to analyse a fragmented production chain is to view production as a set of tasks that are required to produce a single good. Trade, in this view, no longer is the trade in final goods, but, trade in the tasks that are required for the production of that good.
5 There are two approaches to analysing global value chain trade on a global level (Baldwin and Lopez-Gonzales, 2014). Firstly one could use data on trade in intermediate goods. This option relies on the classification using HS codes which describes certain exports as being 'parts' or 'components'. However this approach is unsatisfactory due to two reasons. The HS codes are far from reliable in describing whether exports are actually final or intermediate goods. As can be seen in figure 2 some type of goods (such as car tires) can be used both as an intermediate good and as a final good. HS-codes are not able to clearly label car tires as either being spare tire or as an intermediate good (Baldwin and Lopez-Gonzales, 2014). Moreover, even if these labels were reliable, the data only shows information on the quantity exported and do not describe the intricate networks of the production process as sketched above. The second approach is to use input-output tables. Input-output tables keep track of how much of the inputs that a sector uses is produced in another sector. Recently many authors have adopted this method (See: Hummels, Ishii and Yi, 2001; Johnson and Noguera, 2012; Timmer et al., 2013; Koopman et al. 2014; Baldwin and Lopez-Gonzales, 2014).
Task 1 + Task 2 Households
Country A Country C
The First Unbundling
The Second Unbundling
Country A Country B Country C
Task 1 Task 2 Households
Figure 1: The Two Unbubdlings Before The First Unbundling
Task 1 + Task 2 Households
6 The advantage of using this methodology is that one does have clear statistics on how much input is actually used by a sector and the methodology of input-output analysis (Leontief, 1936) is designed to account explicitly for linkages between sectors. The usage of input-output tables however has as a disadvantage that it uses more aggregated information (the information provided is only available for a limited set of broadly defined sectors) and it needs a number of simplifying assumptions. For example in the methodology of input output analysis one has to assume constant returns to scale and also that there is no substitution in the production process at a given time period2.
Although there are different studies of global value chains in the context of input-output analysis the authors often differ in their approaches. Hummels, Ishii and Yi (2001) use a calculation of vertical specialization that represents the amount of imports that are embodied in the exports of a country. As Baldwin and Lopez-Gonzales (2014) illustrate, this representation is somewhat limited in the analysis of global value chains. The reason for this is that Hummels, Ishii and Yi calculate the amount of imports required to export a certain amount of goods for a number of countries. However the share of imports required to export does not fully capture global value chains. Hummels, Ishii and Yi(2001) use national input-output tables for their estimations. Therefore they cannot account for the
2 This point is often confused with the fact that the methodology of input-output analysis does not allow for substitution
over time. This however is not case since coefficients of the input-output analysis change over time and hence in this sense there is substitution. However substitution does not take place in a single time period.
Spider
Spider
Snake Snake
Task Task Task Task
Task Task Task Task
Households Assembly
Figure 2: Global value chain: Snakes and Spiders
7 origin of the imported products. The imported products themselves may be produced using imports from other countries. Hence, although Hummels, Ishii and Yi might be able to estimate a country’s participation in global value chains, they cannot analyse the global value chains themselves.
Baldwin and Lopez-Gonzales (2014) use a number of indicators of global value chain trade: the imports to produce, which is foreign intermediate goods for the domestic production process. Imports to exports we have already discussed this is the amount of imports embodied in exports of a country. Finally Baldwin and Lopez-Gonzales look at the reimports and reexports3. Reimporting is the process in which goods are initially exported as intermediate goods but imported back. Reexporting is the process in which goods are initially imported as intermediate goods but are later exported to the same country as it was imported from.
The measures that are more closely related to our own, however, are those of Koopman et al. (2014) and Timmer et al. (2013). Both studies use the value added coefficients (more on this in the methodology section) to calculate the developments of global value chain trade. The reason we prefer measures based on value added is that it directly measures the involvement of countries in the global value chain. For example the iPad and iPhone devices are assembled in factories in China. But the Chinese value added that is generated only constitutes about $10 of the price of an iPad or iPhone which are sold for $229 to $275 (Kraemer, Linden and Dedrick, 2011). To analyse global value chains we will trace the actual value added that is involved in the production process.
2.3. Specialization and fragmentation
Theoretically, the effect of the rise of global value chain trade has the same effect on specialization as theorized in the neoclassical model. In the first unbundling there exists a difference in the relative price of commodities between two countries, caused by comparative advantages either due to technological difference or from differences in relative factor endowment. Now if it is the case that transport costs between countries exceeds the difference in the relative prices of the goods then it will not be profitable to trade. If transport costs however go down due to innovations in transport technology, trading becomes profitable. According to the H-O model differences exist due to relative difference in factor abundance. Countries will start exporting (and hence specializing in the production of) those goods that require more inputs in which the country is relatively well endowed. This shows the way in which neoclassical trade models explain that countries specialize in production when trade barriers decrease.
3 The concept of reimport and reexport should not be confused with transit trade. The difference lies in the fact that
8 The Hecksher-Ohlin model can also be used to explain specialization in the setting of global value chains (Deardorff, 2001, Timmer et al., 2013). In this case it is assumed it is not transport costs that impedes trade, but rather something we will call coordination costs. For our model suppose there are two firms in two countries, Country A and Country B, producing the same good. The production of this good requires two tasks, task I is highly dependent on low skilled labour and task II requires a lot of high skilled labour. We assume that factors of production are immobile between countries and that country A is well endowed in high skilled labour and country B is well endowed in low skilled labour. One firm is located in country A and the other firm is located in country B. A firm could export one of its production tasks to the other country however doing so would incur a certain amount of coordination costs. Because of the relative endowment of low and high skilled labour the production of task I is relatively cheaper in country B and task II relatively cheaper in country A. If the coordination costs exceed the difference in the production costs it would not be profitable for the firm to offshore part of its production process. In this case country A and country B both produce tasks I and II. If, due to some technological change, the coordination costs will drop in such a way that it would be profitable to relocate, firms in both countries will offshore their production to minimize costs. In this example each firm will start importing tasks from the other country and we can see instantly how this trade will cause specialization. In this case country A will start producing more task II and country B will start producing more task I.
9 The literature discussed above shows the apparent link between international fragmentation and specialization from the perspective of the H-O model. In this paper we will hypothesise that international fragmentation causes specialization. The reasoning is as follows: because of technological developments, information and communication costs have gone down which made coordination between production tasks cheaper. Because of this development it is now economically feasible to perform different production tasks in different countries. Due to this development, firms seek to lower their costs by importing some tasks from countries that can produce these tasks more efficiently due to differences in the relative factor endowments of the countries. This trade will lead to an increase in international fragmentation. At the same time, international competition will crowd out tasks which are relatively expensive to produce domestically and due to the competitive advantage those tasks that are produced relatively cheaply will be exported to other countries. This effect will lead to an increase in the export specialization of a country. Hence:
H1: An increase in the extent of international fragmentation will lead to an increase in export specialization.
We will be testing this hypothesis by using panel data from the years 1995-2007, using input-output analysis applied to the WIOD-database. In the next section we will expand upon the data and methodology used in the writing of this paper.
3. Data and Methodology
3.1. The basics of input-output analysis
For the analysis of global value chains we heavily rely on multi-regional input-output tables. In figure 3 we give a stylized representation of a multi-regional table for a world that only consists of three countries. The WIOD-database includes 40 countries plus an estimate for the rest of the world (ROW). 𝑍 is the matrix containing economic transactions between sectors. The elements 𝑍𝑖𝑗𝑟𝑠 give the amount of deliveries of sector 𝑖 in country 𝑟 to sector 𝑗 in country 𝑠. 𝑓 is a vector containing final demand, the elements 𝑓𝑖𝑟4 represent the final demand for goods and services produced by
Figure 4: A stylized representation of a multi-regional Input-output table
4 Ordinarily 𝑓 is in the form a matrix with the elements 𝑓
𝑖𝑟𝑠𝑐 meaning the amount of demand for final products produced
10
County A Country B Country C Final
Demand Total Output Country A 𝑍11 𝑍12 𝑍13 𝑓1 𝑥1 Country B 𝑍21 𝑍22 𝑍23 𝑓2 𝑥2 Country C 𝑍31 𝑍32 𝑍33 𝑓3 𝑥3 Value Added 𝑉1′ 𝑉2′ 𝑉3′ Total Output 𝑋1′ 𝑋2′ 𝑋3′
sector 𝑖 in country 𝑟. The vector 𝑥 is a vector with the total amount of economic transactions which include. The elements 𝑥𝑖𝑟 gives the total output of sector 𝑖 in country 𝑟. Finally the input-output table also contains the vector 𝑉 which is a vector of value added. It contains the elements 𝑉𝑖𝑟which gives the value added generated in sector 𝑖 in country 𝑟.
We can also derive the relative inputs from matrix 𝑍, this gives us matrix 𝐴 which consists of the elements 𝑎𝑖𝑗𝑟𝑠, these elements give the inputs from sector 𝑖 in country 𝑟 per dollar of output produced by sector 𝑗 in country 𝑠. These give insight into the relative amount of inputs that are involved in the total output. The matrix is 𝐴 is obtained as 𝐴 = 𝑍𝑥̂−1. With 𝑥 = 𝐴𝑥 + 𝑓 it also follows that 𝑥 = (𝐼 − 𝐴)−1𝑓. The matrix 𝑀 = (𝐼 − 𝐴)1is the so-called Leontief inverse. It captures the indirect and direct linkages that are involved in the production process. The elements of 𝑀 are 𝑚𝑖𝑗𝑟𝑠 and give the total amount of output required in sector 𝑖 in country 𝑟 to satisfy one dollar of final demand in sector 𝑗 in country 𝑠. To illustrate what the Leontief inverse, let 𝑓̃ be a vector of a desired amount of final demand. To produce 𝑓̃ one requires a certain amount of intermediate inputs 𝐴𝑓̃ and these intermediate inputs themselves also need to be produced and require 𝐴2𝑓̃ of inputs. Again these intermediate goods need to be produced which requires 𝐴3𝑓̃. The calculation of the total amount of output required to produce 𝑓̃ is equal to ∑∞ 𝐴𝑘𝑓 ̃
𝑘=0 and this equals (𝐼 − 𝐴)−1𝑓̃. The Leontief inverse, hence, can also be interpreted simultaneously capturing the round by round output required to produce a certain amount of final demand.
3.2. Value added in input-output analysis
11 simply calculate the value added by 𝑉 = 𝑣̂(𝐼 − 𝐴)−1𝑓. This however only gives a vector containing information about the total value added in a certain sector in a certain country. For the intends and purposes of our paper we are interested in the decomposition of this value added into its constituent parts, i.e. how much value added is generated in the Spanish agriculture sector embodied in the Dutch food manufacturing sector, for example. This captures the linkages in a global value chain. The decomposition can be given by 𝑉 = 𝑣̂(𝐼 − 𝐴)−1𝑓̂ . This gives a matrix with the same dimensions as 𝐴. The elements 𝑉𝑖𝑗𝑟𝑠 give us the amount of value added embodied in sector 𝑖 in country 𝑟 due to final demand for products of sector 𝑗 in country 𝑠.
3.3. Adding skill levels
Above we have described how we make a decomposition of value added for the different sectors described in the WIOD. However, as has become clear from the literature section, the existence of specialization as we have hypothesised exists on the level of different tasks, not on the level of sectors. To account for the tasks we will make a further decomposition of value added into three different labour-skill levels: high-skilled, medium-skilled and low-skilled. We do so by using three different vectors, respectively:𝐻𝑆, 𝑀𝑆 and 𝐿𝑆, using the example of 𝐻𝑆 the vector 𝐻𝑆 has the elements 𝐻𝑆𝑖𝑟 which is the share of the compensation of high-skilled workers in total labour compensation in sector 𝑖 in country 𝑟 with 𝐻𝑆𝑖𝑟+ 𝑀𝑆
𝑖𝑟+ 𝐿𝑆𝑖𝑟 = 1. Based on this information we derive three different vectors of value added coefficients, using the example for high-skilled labour we attain the vector for high-skilled value added coefficients by 𝑣′𝐻𝑆 = 𝑣′𝐻𝑆̂ . We use these three different value added vectors to calculate three different matrices for V representing the different skill types, once again using the example for high-skilled labour we do so by 𝑉𝐻𝑆 = 𝑣̂ (𝐼 − 𝐴)𝐻𝑆 −1𝑓̂. We use 𝑉
𝐻𝑆,𝑉𝑀𝑆 and 𝑉𝐿𝑆 to redistribute the value added generated in sector 𝑖 and estimate the value added in sector 𝑖 attributed to high-skilled, medium-skilled and low-skilled tasks. These matrices are combined to create a larger matrix 𝑉𝑡𝑜𝑡, figure 4 gives a stylized representation of a world consisting of two countries and two sectors.
12 Figure 4: Stylized representation of Matrix 𝑽𝒕𝒐𝒕
Country A Country B
Sector I Sector II Sector I Sector II Country A Sector I HS MS LS Sector II HS MS LS Country B Sector I HS MS LS Sector II HS MS LS
3.4. Export Specialization and the extent of fragmentation.
For determining the relationship between the extent of fragmentation and export specialization we need variables that are able to approximate both terms. For export specialization we have used a measure based on the works of Van der Linden and Oosterhaven (2001). The measure of export specialization needs to be able to compare the amount of specialization in a country, as compared to other countries. On the sector level this specialization is given by the formula:
𝐄𝐒𝐢𝐫= ∑ ∑ 𝐕𝐢𝐣 𝐫𝐬 𝐤 𝐬≠𝐫 𝐧 𝐣=𝟏 ∑ ∑ ∑𝐤 𝐕𝐢𝐣𝐫𝐬 𝐬≠𝐫 𝐧 𝐣=𝟏 𝟑𝐧 𝐢=𝟏 − ∑ ∑ ∑ 𝐕𝐢𝐣 𝐫𝐬 𝐤 𝐬≠𝐫 𝐤 𝐫=𝟏 𝐧 𝐣 ∑ ∑ ∑ ∑𝐤 𝐕𝐢𝐣𝐫𝐬 𝐬≠𝐫 𝐤 𝐫=𝟏 𝐧 𝐣=𝟏 𝟑𝐧 𝐢=𝟏
13 𝐄𝐒𝐫= ∑|𝑬𝑺 𝒊 𝒓| ∗𝟏 𝟐 𝟑𝐧 𝐢=𝟏
The variable is based on the absolute values of the sector level specialization, implying that both positive and negative values of sector level specialization contribute to the GVC specialization of a country. The way to interpret this is that the relative specialization in one sector comes at the cost of the share of the other sector. The variable takes a value between 1 and 0, 1 being a country that has completely specialized 0 being a country that has no specialization at all.
To estimate the extent of international fragmentation we estimate the share of GDP that is embodied in foreign final demand. We will call this measure GVCP which stands for Global Value Chain Participation. 𝐆𝐕𝐂𝐏𝐫 = ∑𝐧𝐢=𝟏∑𝐧𝐣=𝟏∑𝐤𝐬≠𝐫𝐕𝐢𝐣𝐫𝐬 ∑ ∑ ∑𝐤 𝐕𝐢𝐣𝐫𝐬 𝐬=𝟏 𝐧 𝐣=𝟏 𝐧 𝐢=𝟏
This term is the share of value added that is embodied in foreign final demand. The amount of value added generated by exporting intermediate goods is given in the numerator, the total value added is given in the denominator.
Our hypothesis is that the extent of international fragmentation causes specialization. It therefore becomes clear that our intent is to regress ES on GVCP (amongst other things), but, before we present our regression equation, we would first like to discuss some potential problems that are common to econometrical analysis in this context.
3.5. Endogeneity and Stationarity
14 regression and the first stage of a 2SLS IV regression. If a variable x is endogenous the error from the first stage regression correlates with the dependent variable of the original regression. This test should be able to identify if either our model suffers from a non-country specific omitted variable bias, or reversed causality.
We use the Hausman test to test for the endogeneity of the first difference of GVCP using the instruments first difference of the share of exports of final goods on GDP and first difference of the share of exports of final goods on total exports. Both variables are assumed to exogenous because they deal with the quantities of exports in general without saying anything about the contents of this trade, moreover they say something about the quantities of final goods and hence cannot tell anything about the type of intermediate goods exported. Both variables are reasonably effective in estimating GVCP with an F-statistic of (F=14.39). Results from this test do not show a significant correlation with the error term of the first stage regression and export specialization in the original regression with a p-value of (p=0.497), hence evidence shows no concern for the adverse effects of endogeneity either being from reversed causality or a non-country specific omitted variable.
Another common problem with panel data is the existence of non-stationary variables. Because panel data involves time series information it can be conceivable that two variables might only be spuriously related, i.e. two unrelated variables might be correlated because of the simple fact that they develop similarly over time. Since we expect the variables ES and GVCP to increase over time this is a very distinct possibility. To test for the stationarity of our variables we have used a modified Dickey-Fuller test for panel data (based on panel means). The Dickey-Fuller test shows that ES and GVCP are indeed non-stationary but the variable ES is stationary around a trend. To overcome the problem of stationarity we will be regress on the first difference of our variables and because our dependent variable is stable around a trend we will also include a time trend variable. This gives us the following regression estimation:
𝚫𝑬𝑺𝒓,𝒕 = 𝜶𝒓+ 𝜹𝒕 + 𝜷𝟏𝚫𝑮𝑽𝑪𝑷𝒓,𝒕+ 𝛄𝟏𝑪𝑭𝑰𝐫,𝐭+ 𝛄𝟐 𝚫𝒍𝒏𝐆𝐃𝐏𝐫,𝐭+ 𝐞𝐫,𝐭
15 deflators5. This variable is included because we assume that countries that are smaller do not have a big enough market size to develop sufficient economies of scale for multiple many sectors, larger markets can develop economies of scale for more sectors because of their own size. Hence we assume that there is a negative relationship between Export specialization and the size of the economy. Because of the size difference of GDP for different countries and hence the potential for outliers we take the natural logarithm of GDP.
3.6. Data:
The bulk of our data is obtained from the WIOD, the WIOD provides a harmonized multi-regional input-output database for 40 countries plus an approximation for the rest of the world (ROW) and 35 sectors6 (See Timmer (ed), 2012; and Dietzenbacher et al, 2013). The WIOD contains (at the time of writing) both national and multiregional input-output tables from 1995 to 2011 and also includes socio-economic and environmental satellite accounts. For the purposes of this project we will primarily be using the World Input Output Tables (WIOT). We use panel data based on the WIOT tables for the years 1995-2007, and do not include data after this date because of possible contagion of effects due to the recent financial crisis.
Many countries in this database are lacking data on the sector: private households with employed persons. Because of this lack of data our ES estimator will be inflated for these countries. As discussed ES measures the relative differences in GVC contributions, it counts both positive and negative differences towards the specialization of the country. Since these countries are seen as not to producing anything in this sector, it will artificially increase our ES variable. Therefore we will exclude this sector from our analysis. Furthermore we have a limited set of 57 observations for which there are no available data for a specific sector in a country in a certain year. As discussed above the lack of data will artificially inflate the ES measure, therefore we will treat these observations as 𝐸𝑆𝑖𝑟 = 0. Appendix 1 includes a comprehensive list of these observations.
We also will use the socio-economic accounts to create the control variables pertaining to capital stock formations. For this variable we will use the gross capital formation data from the socio-economic accounts. This data is expressed in millions of national currency. The data will be converted into millions of dollars using the same exchange rates the authors used to convert the data of the WIOT. The exchange rate data is also available at the WIOD website. The WIOD exchange rates (year averages) were collected from the International Financial Statistics database of the International
5 We use the USA deflators because the WIOD data is expressed in US dollars. Due to the Balassa-Samuelson we
assume that any differences between the relative price levels between countries is captured by the exchange rate, hence the only difference in price levels that remain are those over time in the USA.
16 Monetary Fund (IMF). We will also use the data on the shares of total labour compensation of different skill groups (𝐻𝑆, 𝑀𝑆 and 𝐿𝑆) from the socio-economic accounts of the WIOD.
One final issue is that there are no socio-economic accounts for the ROW estimate hence we need to estimate these numbers. We have estimated these numbers on the basis of the unweighted average of the four ‘most similar’ countries. We have identified this by comparing the differences in the 13-year average values for GVCP, ES (without skill groups) and GDP growth rate. On the basis of this data we found that the countries: “Korea, the UK, Slovenia and Canada” are most similar to the ROW. Hence the data based on the socio-economic account for the ROW will be the unweighted average of these four countries.
4. Results
The presentation of the results will be divided into two parts. The first part contains the results from the raw data from the input-output analysis. This data provides with a general overview of the development of our variables and some of the underlying causes of this development. In the second part we will present our econometric analysis of our theoretical model.
4.1. Input-output analysis
17 Table 1: 13-year average values Table 2: Most and Least specialization and fragmentation for GVCP and ES country GVCP ES aus 0.138 0.163 aut 0.199 0.107 bel 0.256 0.110 bgr 0.203 0.197 bra 0.072 0.138 can 0.195 0.136 chn 0.107 0.159 cyp 0.090 0.110 cze 0.233 0.116 deu 0.149 0.114 dnk 0.162 0.059 esp 0.092 0.138 est 0.247 0.150 fin 0.215 0.170 fra 0.107 0.083 gbr 0.138 0.064 grc 0.062 0.131 hun 0.201 0.093 idn 0.187 0.275 ind 0.072 0.167 irl 0.292 0.190 ita 0.103 0.132 jpn 0.066 0.102 kor 0.153 0.126 ltu 0.179 0.167 lux 0.437 0.128 lva 0.207 0.165 mex 0.111 0.107 mlt 0.229 0.157 nld 0.221 0.069 pol 0.136 0.093 prt 0.103 0.181 rou 0.136 0.174 rus 0.245 0.131 svk 0.243 0.129 svn 0.188 0.130 swe 0.202 0.115 tur 0.059 0.197 twn 0.216 0.139 usa 0.055 0.090 row 0.160 0.195 Mean 0.167 0.136
Most fragmented countries Most specialized countries
Luxembourg Indonesia
Ireland Bulgaria
Belgium Turkey
Estonia Rest of the world
Russia Ireland
Least Fragmented countries Least specialized countries
USA Denmark
Turkey Great Britain
Greece The Netherlands
Japan France
Brazil The United States of
18 In figure 5 we can see the development of both ES and GVCP over time, we have included both a graph showing us the general development of both variables and the development of the first differences. The data as presented in figure 5 seems to be indicative of a correlation between ES and GVCP. The first panel seems to indicate a slow upwards trend of both and, both, seemingly having the same peaks and falls. This is shown more clearly in the bottom panel where we clearly see that both variables have very similar movements.
We have also provided a similar figure for different important economies. The data on the development of GVCP and ES for these countries is presented in figure 6. On the left panel we can see the general development of GVCP and ES, while on the right panel we can see the movement of the first difference of GVCP and ES. We created graphs for four different regions: EU-277, the USA, Japan and China. From this data we can see some additional indication that there is indeed a positive link between the GVCP and ES. This relationship seems to be most apparent for the EU-27 countries and from Japan and less so for China and the USA. For China we see that while GVCP is trending upwards ES is trending downwards, one possible explanation for the downward trend of China might be due to the economic progress of China. We assume that GDP growth has a negative impact on specialization, this is due to the ability of countries to achieve economies of scale, since China is one of fastest developing countries the impact of fragmentation on export specialization might be undone by it fast economic growth. For the USA we see some similar movements but the relationship does not appear to be that big, this might be due to the fact that due to its size it generally does not participate in global value chains that much relative to its income.
7 Although it is possible to calculate the actual GVCP and ES for the EU-27 area as a whole, this data is an aggregate
19 Figure 5: Word Weighted Average for ES and GVCP.
-0.015 -0.01 -0.005 0 0.005 0.01 0.015 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007
Development of World Weighted Average of ES and GVCP
ΔGVCP ΔES 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007
Development of world weighted average of ES and GVCP
20 Figure 6: Development of ES and GVCP for selected regions
22 We also present some data on the specialization per sector of a couple of important countries for a couple of important sectors. We present this with the 13 year average specialization of in certain sectors. The sector specialization is given by 𝐸𝑆𝑖,𝑡𝑟 and we get the average by 𝐸𝑆̅̅̅̅̅ = ∑ 𝐸𝑆𝑖𝑟 𝑡 𝑖,𝑡𝑟 /13. The data from this is represented in figure 7. Due to the abundance of data we have limited ourselves to present data for: Brazil, China, Germany, France, the UK, India, Indonesia, Japan, Korea, Mexico, Russia and the USA and provide information for three sectors. These sectors are: agriculture, hunting, forestry and fishing, electrical and optical equipment, and renting of machinery and equipment and other business activities. We have chosen for these sectors because they are big sectors, and, are to some extent representative for agriculture, manufacturing and services.
Looking at the data in figure 7 it appears that the data is in accordance with the intuition that the more developed economies are more specialized in the production of high-skilled activities. Looking at the agriculture, hunting, forestry and fishing sector we see that the developing countries are more specialized in particularly the low-skilled end of it. This is as expected because we assume that developing country has an abundance in low skilled workers and we assume that this sector requires relatively much low-skilled labour. Looking at the electrical and optical equipment sector we can see that the developed Asian countries Korea and Japan are particularly specialized in the production in this sector. We also see that these countries particularly contribute in the form of high and medium-skilled labour. As we would expect we also find that China is specialized in the production in this sector, however this particularly in the form of low skilled labour. This fits the picture of China we have as the place where manufacturers like Apple go to do their assembly work. There is also some specialization in the USA this only takes the form of in the specialization in high-skilled labour. This fits the picture of the USA as an HQ-economy providing the high-high-skilled upstream activities of a global value chain. If we look at the renting of machinery and equipment and other business activities sector, we can see specialization in particular for ‘western’ countries. This is consistent with the picture that western countries are deindustrializing and specializing in services. Once again we see specializing particularly in the high skilled tasks in this particular sector. Another interesting aspect on the data is that European countries do not only specialize in the high-skill task of this sector but also the medium and low-skill. This can partly be explained by the fact that tasks are clustered, i.e. high-skilled tasks also require the support of medium and low skilled tasks. Another contributing factor is the simplifying assumption that all three skill matrices require the same input-coefficients. Hence, ceteris paribus, a higher export of intermediate goods in a sector as compared to other countries affects all skill-levels.
We also provide some information about the growth in ES. To do we look at the thirteen year average of the first difference of the same selected countries and sectors. We calculate this by Δ𝐸𝑆̅̅̅̅̅̅̅ =𝑖𝑟 (𝐸𝑆𝑖,2007𝑟 − 𝐸𝑆
23 Figure 7: Average specialization for selected countries and sectors
-0.12 -0.08 -0.04 0 0.04 0.08 0.12
bra chn deu fra gbr grc idn ind jpn kor mex rus usa
13-year average specialization for Electrical and Optical Equipment
high skill: Electrical and Optical Equipment medium skill: Electrical and Optical Equipment low skill: Electrical and Optical Equipment
-0.12 -0.08 -0.04 0 0.04 0.08 0.12
bra chn deu fra gbr idn ind jpn kor mex rus usa
13- year average for Renting of M&Eq and other Business activies
high skill: Renting of M&Eq and Other Business Activities medium skill: Renting of M&Eq and Other Business Activities low skill: Renting of M&Eq and Other Business Activities -0.12 -0.08 -0.04 0 0.04 0.08 0.12
bra chn deu fra gbr idn ind jpn kor mex rus usa
13-year average specialization for Agriculture, Hunting, Forestry and Fishing
24 figure many of the developing countries are moving away from exporting agricultural products, this is a sign of a development. When economies develop they move away from agriculture and turn to more profitable sectors in which they can produce more efficiently. If we turn to the electrical and optical equipment sector, we see the growth in the developing countries can in particularly be attributed to medium-skilled labour. We see a particularly interesting increase in the Korean high-skilled labour specialization. An explanation for this might be the growth of areas as Daejeon, which is often regarded as a ‘hub’ of silicon valley, the data seems to corroborate the intuition that Korea now serves as an area to which much of the High-skilled technological jobs are offshored to. Turning to renting of machinery and equipment and other business activities we can see can see a particularly interesting rise for both the UK and India. India is often seen as a place to offshore services activities and the data seems to corroborate this, the interesting aspect of this is that it appears to be the case that it is in particular the high-skilled labour which is being off-shored.
25 Figure 8: Average difference of specialization of selected countries and sectors
-0.005 -0.003 -0.001 0.001 0.003 0.005
bra chn deu fra gbr idn ind jpn kor mex rus usa
13-year average of the difference in Agriculture, Hunting, Forestry and Fishing
high skill: Agriculture, Hunting, Forestry and Fishing medium skill: Agriculture, Hunting, Forestry and Fishing low skill: Agriculture, Hunting, Forestry and Fishing
-0.005 -0.003 -0.001 0.001 0.003 0.005
bra chn deu fra gbr idn ind jpn kor mex rus usa
13-year average of difference in Electrical and Opitcal equipment
high skill: Electrical and Optical Equipment medium skill: Electrical and Optical Equipment low skill: Electrical and Optical Equipment
-0.012 -0.006 0 0.006 0.012
bra chn deu fra gbr idn ind jpn kor mex rus usa
13-year average of the difference of rent in M&Eq and other Business activies
26 4.2. Econometric results:
The raw data from the input-output analysis has shown us some indication that there might be a causal link between the extent of fragmentation and specialization. Before presenting the analysis of the regression we provide a correlation matrix to determine if our data suffers from a multicollinearity problem. The results of which are presented in table 3.
Table 3: Correlation Matrix
𝜟𝑬𝑺𝒓,𝒕 𝜟𝑮𝑽𝑪𝑷𝒓,𝒕 𝜟𝒍𝒏𝑮𝑫𝑷𝒓,𝒕 𝑪𝑭𝑰𝒓,𝒕
𝜟𝑬𝑺𝒓,𝒕 1
𝜟𝑮𝑽𝑪𝑷𝒓,𝒕 0.223 1
𝜟𝒍𝒏𝑮𝑫𝑷𝒓,𝒕 -0.155 -0.401 1
𝑪𝑭𝑰𝒓,𝒕 -0.098 -0.003 0.148 1
As we can see from the correlation matrix no pair of independent variables has a correlation more than 0.8, which, generally speaking, is chosen as the cut-off point for problems of multicollinearity. Therefore it does not appear that multicollinearity would be a problem for our model. Moreover we have tested for the possibility of outliers using Grubbs test for outliers. Grubbs test uses the Z values to test for the possible existence of outliers in a dataset. We have performed Grubbs test on the first differences of ES and the first differences of GVCP. With a cut-off Z-score of 3.86 for GVCP for a five percent significance level, we have identified four possible outliers, one in the first differences of export specialization and three in the first differences of GVCP. Because of the relative few instances that are found and because we expect that much of these biases will be captured by the fixed effect estimators we have not excluded these observation from our analysis. However we have ran a regression with a diminished sample as robustness test later in the paper (more on this later on).
27 GVCP in this specification. The addition of the time trend and a lagged variable de-trends ES. We have included this specification to test the robustness of our results. We use it to test whether the results are still significant under a different specification. The magnitude of the coefficients are different since the dependent variable is on a different scale. Also regression (1.3) should have a much greater R2 due to the inclusion of a lagged term, we have already established that ES is trend stationary hence the lagged term should be extremely correlated with the dependent variable.
Table 4: Main set of Regressions
Dependent variable: first difference of Export Specialization (1.1)FE: Control Only (1.2)Fe: Main
regression (1.3)FE: Alternative Speciation 𝑪𝒐𝒏𝒔𝒕𝒂𝒏𝒕 0.0006 (0.0038) 0.0005 (0.0038) 0.0367 *** (0.0054) 𝑬𝑺𝒓,𝒕−𝟏 0.7045 *** (0.0336) 𝒀𝒆𝒂𝒓𝒕 0.0006 *** (0.0001) 0.0005 *** (0.0001) 0.0007 *** (0.0001) 𝚫𝑮𝑽𝑪𝑷𝒓,𝒕 0.0961 *** (0.0349) 0.0837 *** (0.0323) 𝚫𝑮𝑫𝑷𝒓,𝒕 -0.0220 *** (0.0043) -0.0152 *** (0.0049) -0.0133 *** (0.0046) 𝑪𝑭𝑰𝒓,𝒕 -0.0156 (0.0150) -0.0145 (0.0149) -0.0031 (0.0139) 𝑹𝟐 0.0712 (0.0680) 0.0878 (0.0835) 0.9586 (0.5424) * = significant at 10% ** = significant at 5% *** = significant at 1%
Standard error between parentheses and for R2 the parentheses represent the ‘R2 within’8
The results of (1.1) indicate, as we have expected, that GDP has a negative influence on export specialization. We hypothesised this would be the case due to the ability to develop economies of scale. Interestingly however, CFI does not appear to have a significant influence on export specialization. This is somewhat surprising since we hypothesised that if a country has intensive capital formation, the sectors in which capital is formed more intensely would be more competitive and, therefore, show a stronger specialization.
The main results are presented in regression (1.2). In this regression we have added the GVCP variable which approximates the extent of fragmentation. Based on the changes in the R2 it appears to be the case that adding the variable makes the model more powerful. Also the R2 within of (1.2) is bigger than the R2 of regression (1.1) indicating that non-dummy variables are better able to predict the variance in our data. Interestingly however the F-statistic in regression (1.1) (F=10.89) is higher than that of regression (1.2) (F=10.19). The GVCP variable as we have expected is both positive and
28 significant with a coefficient of 0.096. Regression (1.2) seems to corroborate our hypothesis. The results of regression (1.2) are further corroborated by the results of regression (1.3), the alternative specification of our model gives similar results. There are some differences in the coefficients but this is due to the differences of the scale of the dependent variable. Also the R2 is much greater, but this is as expected and can be attributed by the effect of including the lagged ES variable.
To test the robustness of our results we have also run a number of alternative regression. We can distinguish these into four categories. The first category robustness checks are cross-sectional robustness checks in order identify possible biases in the cross-section of our data. In the second category of robustness check if the relation between export specialization and international fragmentation exists on a different level of analysis. In these robustness checks we switch from a national level of analysis to a task-level of analysis. We also use the preform some cross-sectional robustness checks in this category. The third category of robustness checks are time-series robustness checks, we do so in order to observe if there are any possible biases in the time-series of our data. The fourth category are miscellaneous robustness checks.
The first set of robustness checks are presented in table 5. We have performed multiple regressions on the basis of an alternative cross-sectional selection. Based on the raw data from the input-output analysis we have identified a possible concern for an EU-bias. The correlation between appears to be the strongest for EU-countries and since the EU make up 27 of the 41 economies, this might be a legitimate concern for us. To test for this bias we ran regressions only for EU countries in regression (2.1) (324 observations) and for non-EU countries in regression (2.2) (168 observations). We have also used alternative regressions based on specification of different geographical regions and for emerging economies. In regression (2.3) we regress only for East and South-East Asian countries. These are China, India, Indonesia, Japan, Korea and Taiwan (72 observations). In regression (2.4) we are testing for a possible bias against emerging economies, we do so by looking at the BRIIMT countries, these are Brazil, Russia, India, Indonesia, Mexico and Turkey (60 observations).
29 Table 5: Cross-sectional Robustness checks
Dependent variable: first difference of Export Specialization (2.1) EU Only (2.2) Non-EU Only (2.4) East and South-East Asian Only (2.4) BRIIMT only 𝑪𝒐𝒏𝒔𝒕𝒂𝒏𝒕 0.0013 (0.0039) -0.0012 (0.0089) -0.0089 (0.0141) 0.0161 (0.0184) 𝒀𝒆𝒂𝒓𝒕 0.0008 *** (0.0002) 0.0003 (0.0003) -0.0005 (0.0005) -0.0004 (0.0005) 𝚫𝑮𝑽𝑪𝑷𝒓,𝒕 0.1366 *** (0.0390) 0.0916 (0.0683) 0.1815 * (0.1060) 0.1264 (0.1377) 𝚫𝑮𝑫𝑷𝒓,𝒕 -0.0365 *** (0.0064) -0.0020 (0.0087) 0.0008 (0.0136) -0.0124 (0.0154) 𝑪𝑭𝑰𝒓,𝒕 -0.0223 (0.0154) -0.0053 (0.0350) 0.0382 (0.0539) -0.0431 (0.0516) 𝑹𝟐 0.1482 (0.1852) 0.0433 (0.0341) 0.0275 (0.1172) 0.1662 (0.2008) * = significant at 10% ** = significant at 5% *** = significant at 1%
Standard error between parentheses and for R2 the parentheses represent the ‘R2 within’
correlation between export specialization and international fragmentation outside the European Union. We have two possible explanations for the apparent EU bias. Firstly, the EU country set is characterised by smaller open economies, the other countries, with exception of Taiwan tend, are far larger economies and perhaps these countries are therefore also more vertically integrated. Another possible explanation might be that the EU-market is very integrated witch possibly makes fragmentation and therefore specialization more easy between its member states. Another interesting results from this set of regressions is that the only significant result for GDP can be found in the EU-only regression this might be due to the fact that with the exception of Taiwan in the Asian set, the set of EU-countries is the only set that contains small open economies. As we have hypothesised small open economies are more specialized because they cannot achieve economies of scale for many different sectors in their own market.
In the second set of robustness checks we have switched our level of analysis from a national level to the level of tasks. In this level we use the variables we use the variables 𝐸𝑆𝑖,𝑡𝑟 and 𝐺𝑉𝐶𝑃𝑖,𝑡𝑟. Also the dummies of the fixed effect are based on the combination of tasks in a certain sector, i.e. a dummy for high-skilled food processing in the Netherlands etcetera. This effectively increases our observation pool by a factor of 102 (34 sectors with each 3 skill-levels). The formula for 𝐸𝑆𝑖,𝑡𝑟 is already given in the methodology section of the paper. The formula for GVCP is 𝐺𝑉𝐶𝑃𝑖,𝑡𝑟 = ∑𝑛𝑗=1∑𝑘𝑠≠𝑟𝑉𝑖𝑗𝑟𝑠
30 with the exceptions that the coefficient for GVCP should be much smaller, since ES would be much smaller. We also expect to find that GDP is significant though positive. The reason we expect GDP now to be positive is for the same reason we expected it to be negative for our analysis on the national level. The size of economy helps to achieve economies of scale. While on a national level this would have a negative effect since more sector are able to achieve economies of scale. On a national level this would be positive since that specific sector is able to achieve economies of scale. The previous set of regressions helped us identify a bias towards European countries in our dataset, in regression (3.2) and (3.3) we try test whether this bias still exists with an increased sample size. In (3.2) we only use European countries in our dataset (32.955 observations) (3.3) we only use non-European countries in our dataset (17.064).
Table 6: Task-level Robustness Checks
Dependent Variable: the first difference of Export Specialization (task level) (3.1) FE: main
regression
(3.2) FE: EU-only (3.3) FE: non-EU only
𝑪𝒐𝒏𝒔𝒕𝒂𝒏𝒕 0.0000 (0.0001) 0.0000 (0.0001) 0.0000 (0.0002) 𝒀𝒆𝒂𝒓𝒕 0.0000 (0.0000) 0.0000 0.0000 0.0000 0.0000 𝚫𝑮𝑽𝑪𝑷𝒓,𝒊,𝒕 0.0088 *** (0.0002) 0.0084 *** (0.0002) 0.0141 *** (0.0008) 𝚫𝑮𝑫𝑷𝒓,𝒕 0.0006 *** (0.0001) 0.0004 * (0.0002) 0.0123 *** (0.0002) 𝑪𝑭𝑰𝒓,𝒕 0.0001 (0.0004) 0.0001 (0.0005) 0.0000 0.0000 𝑹𝟐 0.0304 (0.0314) 0.0359 (0.0377) 0.0214 (0.0202) * = significant at 10% ** = significant at 5% *** = significant at 1%
Standard error between parentheses and for R2 the parentheses represent the ‘R2 within’
The results for regression (3.1) show a correlation between GVCP and ES. The coefficient is very small but this is as expected. The results are very significant with a t-statistic of (t=38.55). We also see that GDP as expected significant and positive. Regression (3.1) further corroborates our hypothesis on the task-level data. If we compare regression (3.2) and (3.3), at first glance, the opposite seems to be the case as with regression (2.1) and (2.2). Indeed GVCP has a greater coefficient in (3.3). However if we look deeper into the data we see that the results for (3.3) GVCP is far less significant for with a t-statistic of (t=17.96) compared to regression (3.2) (t=34.39). Hence we can say that also say with an increased sample size there seems to be some bias towards European countries, for the correlation appears to be much greater. That being said in contrast to our cross-sectional robustness test we do find a statistically significant value for GVCP in regression (3.3).
31 test whether or not there might be some bias in the time-series dimension. We do not expect different time spans to make a great difference. Figures 5 and 6 do seem to indicate some weaker correlation after 2002 but we think this effect is negligible. Due to the limited time span, we will divide our analysis into two different time spans of roughly the same size the first time span runs from 1995 to 2001 (3.1) the other runs from 2002-2007 (3.2). Both regressions contain 246 observations. The results of this are presented in table 7.
Table 7: Time Series Robustness Checks
Dependent Variable: the first difference of Export Specialization (4.1): 1995-2001 (4.2) 2002-2007 𝑪𝒐𝒏𝒔𝒕𝒂𝒏𝒕 -0.0072 (0.0070) -0.0015 (0.0082) 𝒀𝒆𝒂𝒓𝒕 0.0010 *** (0.0003) -0.0003 (0.0004) 𝚫𝑮𝑽𝑪𝑷𝒓,𝒕 0.1156 ** (0.0509) 0.0839 (0.0540) 𝚫𝑮𝑫𝑷𝒓,𝒕 -0.0121 (0.0073) -0.0207 ** (0.0101) 𝑪𝑭𝑰𝒓,𝒕 0.0070 (0.0267) 0.0333 (0.0320) 𝑹𝟐 0.1145 (0.1323) 0.0025 (0.0536) * = significant at 10% ** = significant at 5% *** = significant at 1% Standard error between brackets and for R2 the brackets represent the ‘R2 within’
The results from regression (3.1) and (3.2) seem to indicate that there might be some difference in the defined time spans for the regression analysis. Between 1995-2001 the correlation between ES and GVCP seems to be greater than between 2001-2007. The differences however do not appear to be very big hence we account these difference as the result of a smaller sample sizes. Very puzzling addition result is that it appears to be the case the GDP only has a significant influence in the time period 2001-2007 and also the time trend is negative through insignificant after this time period. We do not have an explanation for this effect.
32 which has a higher Z score of 3.88, any observation in with a lack of data problem (for in this case 𝐸𝑆𝑖𝑟 for that sector is estimated at zero) and also we have excluded the ROW estimate for we had to estimate a number of variables as the unweighted average of a number of ‘similar countries’. We will present a list of observations we have excluded in appendix 3. Regression (5.3) is based on 437 observations. Finally we also estimate the robustness of the results by using robust standard errors in regression (5.4) to account for the possibility of heterogeneity of our explanatory variables.
Table 6: Miscellaneous robustness checks
Dependent variable: the first difference of export specialization (5.1) Main regression (5.2): Based on VS (5.3) Reduced Sample (5.4) Robust standard errors 𝑪𝒐𝒏𝒔𝒕𝒂𝒏𝒕 0.0005 (0.0038) 0.0007 (0.0038) 0.0021 (0.0037) 0.0005 (0.0044) 𝒀𝒆𝒂𝒓𝒕 0.0005 *** (0.0001) 0.0006 *** (0.0001) 0.0006 *** (0.001) 0.0005 *** (0.0001) 𝚫𝑮𝑽𝑪𝑷𝒓,𝒕 0.0961 *** (0.0349) 0.1160 *** (0.0393) 0.0961 *** (0.0422) 𝚫𝑽𝑺𝒓,𝒕 0.0362 (0.0341) 𝚫𝑮𝑫𝑷𝒓,𝒕 -0.0152 *** (0.0049) -0.0206 *** (0.0045) -0.0188 *** (0.0051) -0.0152 ** (0.0067) 𝑪𝑭𝑰𝒓,𝒕 -0.0145 (0.0149) -0.0159 (0.0150) -0.00214 (0 .0148) -0.0145 (0.0172) 𝑹𝟐 0.0878 (0.0835) 0.0735 (0.0703) 0.0744 (0.0745) 0.0878 (0.0835) * = significant at 10% ** = significant at 5% *** = significant at 1%
Standard error between brackets and for R2 the brackets represent the ‘R2 within’
Although both variables VS and GVCP are intended to measure the same phenomenon: international fragmentation the correlation of the two variables is not as strong as one would suspect with (r=0.341). We also see some significant difference if we were to adopt the VS variable. Although GVCP and VS are roughly of equal size we can see from regression (5.2) that the coefficient of VS is much smaller and statistically insignificant. All the other results seem to be very much in line with regression (5.1), indicating that a different independent variable does not make much of a difference. The results from regression (5.2) seems also to validate our choice of using GVCP as an independent variable.
33 Regression (5.4) uses robust errors according to White’s approach, using robust standard errors does not make the results for GVCP insignificant, GDP on the other hand has become less significant.
4.3. Discussion
The main results of regression (1.2) gives us evidence in favour of a causal relationship between the extent of fragmentation and export specialization. However, further robustness checks seem to cast some doubt on our results. In our cross-sectional analysis, it appears to be the case that there only is a correlation between the two variables for European countries and not for countries outside the European Union. We hypothesised this might be attributed to the economic integration of these countries.
Based on this points we can say that it appears to be the case that there is some positive evidence to conclude that indeed there is a causal link between the extent of fragmentation and export specialization in global value chains. However much more research needs to be done before we are able to confirm our hypothesis. Although we have some positive evidence, we cannot say we have found unambiguous proof in favour our hypothesis, we therefore cannot conclude that we have found a positive causal link between specialization and fragmentation. We see our research as exploratory research for the empirical link between traditional trade theories in the context of modern developments in trade. We hope that future research will continue in similar trend that we have been going on and will improve upon the methodology used in the writing of this paper and that more data will become available giving people the ability to improve upon our methodology.
5. Conclusion
The conclusion will be divided into three parts first we will discuss the limitations of our methodology. Second, we will present the actual conclusion from our analysis. Finally, we will end with some prospects for future research
5.1. Limitations
34 in China and not about the amount of low-skilled labour in the electrical and optical machinery sector in China. Should finely grained data become available, this would strongly improve the quality of our propose methodology.
Secondly, in order to perform the calculation of our input-output analysis, we need to make a great number of simplifying assumptions. For instance we assumed in order to estimate the contributions of value added from the different skill-levels we need to assume that the task use the same input-coefficients. Hence, ceteris paribus, if the intermediate exports in a sector go up, the contributions of the skill-levels to value added go up in proportion to the total labour compensation of that skill-level. In reality, it is very unlikely that this would be the case. Another assumption we need to make is that the production of intermediate goods require the same amount of inputs as the final products, another very unlikely assumption which is typically made in input-output analysis.
Thirdly, at this moment the data of the WIOD only goes back to 1995 and although more recent data is available we have chosen to study up to the year 2007 to avoid any possible contagion of our results from the effects of the financial and economic crisis. Therefore we only have a limited time span of 13 year for which we have conducted our research, to make a more substantive claim about the nature of the relationship between international fragmentation and specialization it would be preferable to include a larger time span in the analysis.
Fourthly, the WIOD only has a limited set of countries and the fast majority of the countries included in the WIOD are European countries. A broader set of countries from more diverse regions would be preferable for the accuracy and to avoid possible biases.
5.2. Conclusion
35 more specialised.
The results from various panel data regressions and robustness test gives us a mixed evidence in favour of our hypothesis: fragmentation leads to specialization. The main results from our regression (1.2) shows a positive link between GVCP and ES whereas further robustness checks show that this link only holds for members of the European Union and does not hold for other countries. All in all we can say that there is some positive evidence pointing to the possible existence of specialization within global value chains, however due to the mixed results, future research is needed to corroborate these results more definitive. Hence we cannot give an unambiguous: yes, as an answer to our research question: Is there a positive link between the extent of international fragmentation and export specialization between the years 1995 and 2007? We, however, have found some supportive evidence concerning this relation and hence we feel it is appropriate to say that there is some corroboration of our hypothesis and therefore some evidence in favour of our research question. But all in all more research is needed.
5.3. Prospects for future research
As discussed in the discussion section of our results there is much room to improve upon the methodology of our research. Much of our research is limited due to the availability of data. We hope that as the future might provide future researchers with more detailed data for their research. Our research has provided us with some indication that Europe might be better suited to reap the benefits of global value chains because of their economic integration. This indication however is very hypothetical and very provisional. It is however, in our opinion, a fruitful topic to explore for future research. The idea that the barriers to vertical specialization are lower between EU-member states is intuitively very appealing.
36 have so little empirical evidence.
Acknowledgements:
